Histogram of Oriented Gradients meet deep learning: A novel multi-task deep network for 2D surgical image semantic...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Current multi-task learning for medical image segmentation requires extensive ground truth annotations for all tasks.
• This limits the scalability and efficiency of existing deep learning models.

Purpose of the Study:

• To develop a novel deep multi-task learning method for medical image segmentation.
• To eliminate the need for ground truth annotations for auxiliary tasks by generating unsupervised pseudo-labels.

Main Methods:

• Proposed a multi-task learning framework that leverages unsupervised pseudo-labels for auxiliary tasks.
• Utilized Histogram of Oriented Gradients (HOGs) for unsupervised pseudo-label generation.
• Integrated pseudo-labels with ground truth semantic segmentation masks to jointly train deep networks (UNet, U2Net).

Main Results:

• The proposed method consistently improved performance over counterpart methods on two medical image segmentation datasets.
• Achieved superior quantitative and qualitative results in medical image segmentation tasks.
• Won the FetReg Endovis Sub-challenge on Semantic Segmentation at MICCAI 2021.

Conclusions:

• The novel unsupervised pseudo-labeling approach effectively enhances multi-task deep learning for medical image segmentation.
• The method offers a more efficient and scalable alternative to traditional multi-task learning frameworks.
• Demonstrated the potential of unsupervised learning in conjunction with hand-crafted features for medical image analysis.